Forecasting emergency trolley drug utilisation using ARMA and SVR with external factors

Abstract

Adequate stock management in emergency trolleys is important to ensure that every process requiring medication specifically in hospitals, runs smoothly in any given situation. Stock management based on the value of means (or average usage) is not adequate to account for unpredictable situations that may result in disruptions in drug utilisation and supply. In this study, an investigation to identify possible factors that correlate with the fluctuation of terbutaline injection drug utilisation used in emergency trolleys, using univariate forecasting methods, machine learning (ML), and hybrid models capable of predicting future usage was undertaken. Based on an experimental dataset, it was found that the mean temperature in Mersing, Johor, has the highest negative correlation with terbutaline injection utilisation, at a correlation coefficient value of -0.27 (p-value = 0.0068). Three univariate models and three univariate models with exogenous variables were constructed and compared. All advanced models show better performance than the naive baseline model that served as a benchmark for model building. The univariate models in the analysis consisted of Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Network (ANN) models. Also, the ML models considered in this study including Support Vector Regression (SVR), used the lagged values of terbutaline injection as its input variables. The Autoregressive Moving Average (ARMA) (4,4) model performed better than the ML models, with a Mean Absolute Error (MAE) of 8.9524 and a Root Mean Square Error (RMSE) of 11.4518 in the validation data set. To incorporate the effects of exogenous variables, significant lags of emergency admission and climate variables were used to construct a predictive model from ARMA (4,4). The hybrid model of ARMA-ANN outperformed all other models, with MAE and RMSE values of 8.8571 and 10.8496 respectively. It can also be summarized that models utilising ANN are far better than SVR models due to a variety of factors, including the type of data input and the optimization techniques used to build the SVR models. Future studies should focus on modelling different types of medication used in emergency trolleys. The application of other ML algorithms and optimisation strategies can also be explored for different patterns of data.